On Sun, Jan 20, 2013 at 3:15 PM, Anastasios Tsiolakidis < [email protected]> wrote:
> Well, > > I don't want to sound like the other old white men here, but you are just > twisting a form of supervised learning. I think in the long run we need to > work out interactive models of disambiguation that include touching and > handling and moving around, so that handling informs seeing as much as > seeing informs handling. I can't see this happening without "features", we > need not have hang ups about geometric shapes, rather we need real world > statistics for the more likely combinations of shapes and colors and > assemblies. I guess this is already a tall order and different from edge > detection etc. > My intention is just to outline an abstract framework. In practice, we will employ a lot of heuristics to make the process efficient. But most vision researchers attempt to find short-cuts and ended up wasting even more time =) Yes, interactivity is required for AGI, but I was only focused on visual recognition. It would be nice to see how my model fits in an AGI architecture, I have not spelled it out, though I have proposed an architecture based on logic + reinforcement learning. Now, to make things worse, an AGI will probably need to operate in real > time in its world, so if it is not pointing a camera to a training set but > rather, say, plays football, it will probably need to actively generate > probabilities for different microverses (stochastic versions of its > immediate environment) and then match them best as it can to the incoming > data stream, rather than idly calculate some Bayesian. If you do kinda > believe in evolution etc then you'll agree that the long lineage of > organisms we are related to could hardly have bothered to recognize cones > with attached triangles, but they sure as hell needed to get out of the way > of sharks. Despite our human obsession with our ability to look at > photographs and pigeonhole them, I have the suspicion those "survival > metrics" never went away, and we're probably lucky they didn't. For > example, the vast majority of cognitive and even physiological systems I > know about, including human vision and vision physiology, are primed as is > well known for change, they are looking out for new data (like you being > able to find your mouse pointer much faster after a little vigorous > movement as opposed to scanning the screen). > > Microverses and stream processing all the way! > Yes, the AGI needs to build cognitive models of the world. The cognitive models in turn *can* generate geometric models. These relations may help us design a better AGI architecture... YKY ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
